This library contains headless versions of a number of commonly used tutoring system environments for training simulated students. There are currently two different tutoring systems that can be loaded: the fraction arithmetic tutor and a multi-column arithmetic tutor.
The core of each tutor is a "Symbolic" variant of each tutor. This class maintains the logic for the tutor, including calls for getting the state description, applying selection, action, inputs (SAIs) which returns feedback about whether the provided sai is correct/incorrect, and which provides a method for requesting a demonstration (a valid Sai for the next step).
The Apprentice Learner Architecture can interface directly with these symbolic variants of the tutor.
Next, there are separate classes that wrap these symbolic tutorrs in AI gym environments that can be loaded by a reinforcement learning algorithm, such as those in the stable baseline library (e.g., the PPO algorithm).
Currently, I am exploring multiple representations for the RL tutor:
-
Operators model: this is the closest to what the Apprentice Learner would use. In particular, the agent perceives the same hot-one coded state features that the AL agent would get. It also has 4 discrete action outputs (multi-discrete space), the first is for the selection (all selectable interface elements in the tutor), the second discrete output is for an operator (copy, add, subtract, multiply, etc., these correspond to prior knowledge operators/functions), the next two outputs correspond to the fields that get passed as input to the operator (e.g., two other fields in the tutor interface that currently have values).
-
Digits model, this has the same input as the Operators model (hot-one features that AL would get), but has a different action output. Instead, it has 4 discrete action outputs (multi-discrete space). The first output is for selection (as above), the second is for a digit in the ones place (0-9), the third is for a digit in the tens place (0-9), and the fourth is for a digit in the hundreds place (0-9). Depending on the tutor the number of digits might be more or less depending on what is necessary to solve the task (e.g., the multi-column arith only has a single digit).
-
Pixel model, This has the same output as the Digits model, but has a different input. Instead, the model gets a black and white, pixel representation of the tutor interface. It is not identical to the human tutor represention, but it is a semi-resonable facimile that includes all the information that exists in the human tutor.
These different representations are registered as AI gym models under the names "FractionArith" and "MultiColumnArith" and version numbers "v0" (for operator model), "v1" (for digits model), and "v2" for pixel model. As an example of how to create an operator model for fraction arithmetic and train a PPO model from the stable baselines package, you would use the following code:
import gym
import tutorenvs
env = make_vec_env('MultiColumnArith-v0', n_envs=8)
model = PPO2(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=2000000)See the code in the sandbox/ folder for examples of how to train different
kinds of agents on these different environments. The
sandbox/run_al_fractions.py code shows how to train an Apprentice agent on
the headless fractions tutor.